How to use data.table fifelse with vectors in the arguments?

Say I have this data.frame

DF <- data.frame(one=c(1, NA, NA, 1, NA, NA), two=c(NA,1,NA, NA, NA,1), 
         three=c(NA,NA, 1, NA, 1,NA))

one    two  three         output
  1     NA    NA             one
 NA      1    NA             two
 NA     NA     1           three
  1     NA    NA             one  
 NA     NA     1           three
 NA      1    NA             two

The columns are mutually exclusive.
I need to generate the output

output=c("one","two","three","one","three", "two")

I've tried to to it with data.table fifelse but it

with(DF,fifelse(one==1, "one", fifelse(two==1,"two", "three", na="three"), 
   na=fifelse(two==1,"two", "three", na="three")))

Error in fifelse(one == 1, "one", fifelse(two == 1, "two", "three", na = "three"),  : 
  Length of 'na' is 6 but must be 1

It seems it doesn't accept a vector on the arguments.

dplyr's if_else works well here.

with(DF,if_else(one==1, "one", if_else(two==1,"two", "three", missing="three"), 
   missing=if_else(two==1,"two", "three", missing="three")))

How can I get the same output with data.table?

Any other simple alternative.
With R base I could use

apply(DF,1, function(x) which(!is.na(x)))

and later replace that numbers with characters.

fifelse isn't the best tool for this, I suggest fcase is easier:

data.table

library(data.table)
as.data.table(DF)[, fcase(one == 1, "one", two == 1, "two", three == 1, "three")]
# [1] "one"   "two"   "three" "one"   "three" "two"  

dplyr

The dplyr analog is case_when:

library(dplyr)
with(DF, case_when(one == 1 ~ "one", two == 1 ~ "two", three == 1 ~ "three"))
# [1] "one"   "two"   "three" "one"   "three" "two"  

base R

Both the data.table and dplyr implementations presume knowing the column names a priori. A base-R method that is agnostic to that:

colnames(DF)[apply(DF, 1, which.max)]
# [1] "one"   "two"   "three" "one"   "three" "two"  

(Incidentally, which.max can also be which.min here, really we're just looking for a non-NA value.)

In this case, if you have other columns that should not be considered, you will need to subset the DF within apply(DF, ...) so that it only looks at the desired columns.

Another data.table alterntive:

for (col in names(DF)) set(DF, which(DF[[col]] == 1), j = "output", value = col)

If you have only one non-NA value each row, you can try max.col

> names(DF)[max.col(!is.na(DF))]
[1] "one"   "two"   "three" "one"   "three" "two"

or col + na.omit (but this might be slow if you are pursuing the speed)

> names(DF)[na.omit(c(t(col(DF) * DF)))]
[1] "one"   "two"   "three" "one"   "three" "two"

Benchmarking

microbenchmark(
    f1 = names(DF)[max.col(!is.na(DF))],
    f2 = names(DF)[na.omit(c(t(col(DF) * DF)))]
)

gives

Unit: microseconds
 expr   min     lq    mean median    uq    max neval
   f1  28.5  51.45  92.343  64.40  91.8 1532.5   100
   f2 300.7 527.65 634.755 595.35 691.5 2405.4   100